Review of Multiple Regression

Assumptions about prior knowledge. This handout attempts to summarize and synthesize
the basics of Multiple Regression that should have been learned in an earlier statistics course. It
is therefore assumed that most of this material is indeed “review” for the reader. (Don’t worry
too much if some items aren’t review; I know that different instructors cover different things,
and many of these topics will be covered again as we go through the semester.) Those wanting
more detail and worked examples should look at my course notes for Grad Stats I. Basic
concepts such as means, standard deviations, correlations, expectations, probability, and
probability distributions are not reviewed. Further, the formulas and discussion generally
assume that the means, standard deviations, and correlations are either already available or can
be easily computed.
In general, I present formulas either because I think they are useful to know, or because I think
they help illustrate key substantive points. Your previous courses may or may not have taken
such a mathematical approach, or may have presented the same material but in a different way.
For many people, formulas can help to make the underlying concepts clearer; if you aren’t one of
them you will probably still be ok. A few key formulas and procedures will be used extensively
throughout the course (particularly those relating to F tests and R2 calculations) and you will get
quite a bit of practice working with them.

Linear regression model

k
Y j = α + β1 X 1 j + β 2 X 2 j + ... + β k X kj + ε j = α + ∑ β i X ij + ε j = E (Y j | X ) + ε j
i =1

β i = partial slope coefficient (also called partial regression coefficient, metric coefficient). It
represents the change in E(Y) associated with a one-unit increase in Xi when all other IVs are
held constant.
α = the intercept. Geometrically, it represents the value of E(Y) where the regression surface (or
plane) crosses the Y axis. Substantively, it is the expected value of Y when all the IVs equal 0.
ε = the deviation of the value Yj from the mean value of the distribution given X. This error
term may be conceived as representing (1) the effects on Y of variables not explicitly included in
the equation, and (2) a residual random element in the dependent variable.

Parameter estimation (Metric Coefficients): In most situations, we are not in a position to

determine the population parameters directly. Instead, we must estimate their values from a
finite sample from the population. The sample regression model is written as
k
Y j = a + b1 X 1 j + b2 X 2 j + ... + bk X kj + e j = a + ∑ bi X ij + e j = Yˆ j + e j
i =1

where a is the sample estimate of α and bk is the sample estimate of βk.

Review of Multiple Regression — Page 1

 Computation of bk
Case Formula(s) Comments
1 IV case s xy Sample covariance of X and Y divided by the
b= variance of X
s x2
General case sy Hand computation can be quite tedious when there
bk = bk′ * are many IVs or a lot of cases. These formulas can
s xk be used if the standardized coefficients, standard
deviations, t values, and standard errors are known.
= t k * sbk

Computation of K Compute the betas first. Then multiply each beta

a (all cases) a = y − ∑ bk xk times the mean of the corresponding X variable and
k =1 sum the results. Subtract from the mean of y.

Question. Suppose bk = 0 for all variables, i.e. none of the IVs have a linear effect on Y. What
is the predicted value of Y?

Standardized coefficients. The IVs and DV can be in an infinite number of metrics.

Income can be measured in dollars, education in years, intelligence in IQ points. This can make
it difficult to compare effects. Hence, some like to “standardize” variables. In effect, a Z-score
transformation is done on each IV and DV. The transformed variables then have a mean of zero
and a variance of 1. Rescaling the variables also rescales the regression coefficients. Formulas
for the standardized coefficients include
1 IV case b′ = ryx In the one IV case, the
standardized coefficient simply
equals the correlation between Y
and X
General Case sx k As this formula shows, it is very
bk′ = bk * easy to go from the metric to the
sy
standardized coefficients. There
is no need to actually compute the
standardized variables and run a
new regression.

We interpret the standardized coefficients as follows: a one standard deviation increase in Xk

results in a b’k standard deviation increase in Y.
Standardized coefficients are somewhat popular because
• Variables are in a common (albeit weird) metric. Hence, it is possible to compare magnitudes
of effects
• They are easier to work with mathematically

Review of Multiple Regression — Page 2

• The metric of many variables is arbitrary and unintuitive anyway. Hence, you might as well
make the scaling standard across variables.
Nevertheless, standardized effects tend to be looked down upon because
• They are not very intuitive
• They can be very misleading; for example, when making comparisons across groups. As we
will see, Duncan argues this point quite forcefully.

The ANOVA Table: Sums of squares, degrees of freedom, mean squares, and F.
Before doing other calculations, it is often useful or necessary to construct the ANOVA
(Analysis of Variance) table. There are four parts to the ANOVA table: sums of squares,
degrees of freedom, mean squares, and the F statistic.
Sums of squares. Sums of squares are actually sums of squared deviations about a mean. For
the ANOVA table, we are interested in the Total sum of squares (SST), the regression sum of
squares (SSR), and the error sum of squares (SSE; also known as the residual sum of squares).

Computation of sums of squares

Case Formula(s)
N
General case:
SST = ∑ ( y j − y ) 2 = SSR + SSE
j =1

N
SSR = ∑ ( y$ j − y ) 2 = SST − SSE
j =1

N N
SSE = ∑ ( y j − y$ j ) 2 = ∑ e j 2 = SST − SSR
j =1 j =1

Question: What do SSE and SSR equal if it is always the case that y j = y$ j , i.e. you make
“perfect” predictions every time? Conversely, what do SSR and SSE equal if it is always the
case that y$ j = y , i.e. for every case the predicted value is the mean of Y?

Review of Multiple Regression — Page 3

Other calculations. The rest of the ANOVA table easily follows (K = # of IVs):
Source SS DF MS F
Regression (or SSR K MSR = F=
explained) SSR/K MSR / MSE
Error (or SSE N-K-1 MSE =
residual) SSE/(N - K - 1)
Total SST N-1 MST =
SST/(N - 1)

An alternative formula for F, which is often useful when the original data are not available (e.g.
when reading someone else’s article) is
R 2 * ( N − K − 1)
F=
(1 − R 2 ) * K
The above formula has several interesting implications, which we will discuss shortly.
Uses of the ANOVA table. As you know (or will see) the information in the ANOVA table
has several uses:
• The F statistic (with df = K, N-K-1) can be used to test the hypothesis that ρ2 = 0 (or
equivalently, that all betas equal 0). In a bivariate regression with a two-tailed alternative
hypothesis, F can test whether β = 0. Also, F (along with N and K) can be used to compute
R2.
• MST = the variance of y, i.e. sy2.
• SSR/SST = R2. Also, SSE/SST = 1 - R2.
• MSE is used to compute the standard error of the estimate (se).
• SSE can be used when testing hypotheses concerning nested models (e.g. are a subset of the
betas equal to 0?)

Multiple R and R2. Multiple R is the correlation between Y and Y$ . It ranges between 0 and 1
(it won’t be negative.) Multiple R2 is the amount of variability in Y that is accounted for
(explained) by the X variables. If there is a perfect linear relationship between Y and the IVs, R2
will equal 1. If there is no linear relationship between Y and the IVs, R2 will equal 0. Note that
R and R2 are the sample estimates of ρ and ρ2.

Review of Multiple Regression — Page 4

Some formulas for R2.
R2 = SSR/SST Explained sum of squares over total sum of
squares, i.e. the ratio of the explained
variability to the total variability.
F *K This can be useful if F, N, and K are known
R2 =
( N − K − 1) + ( F * K )
K
This formula uses the standardized coefficients
R 2 = ∑ bk′ ryk and the zero-order correlations between y and
k =1
the x’s. This (esoteric) formula can be useful
when doing path analysis.
One IV case only: Remember that, in standardized form,
correlations and covariances are the same.
R 2 = b′ 2
2
RYH = RYG
2
+ srk2 See below for definition of H, G, and sr2.
k
These formulas are handy in stepwise
2
RYG k
= RYH
2
− srk2 regression procedures.

Incidentally, R2 is biased upward, particularly in small samples. Therefore, adjusted R2 is

sometimes used. The formula is
⎛ ( N − 1)(1 − R 2 ) ⎞ N −1
Adjusted R = 1 − ⎜
2
⎟ = 1 − (1 − R 2 ) *
⎝ ( N − K − 1) ⎠ N − K −1
Note that, unlike regular R2, Adjusted R2 can actually get smaller as additional variables are
added to the model. As N gets bigger, the difference between R2 and Adjusted R2 gets smaller
and smaller.

Standard error of the estimate. The standard error of the estimate (se) indicates how close
the actual observations fall to the predicted values on the regression line. If ε ∼ N(0, σε2), then
about 68.3% of the observations should fall within ± 1se units of the regression line, 95.4%
should fall within ± 2se units, and 99.7% should fall within ± 3se units. The formula is
SSE
se = = MSE
N − K −1
Standard errors. bk is a point estimate of βk. Because of sampling variability, this estimate
may be too high or too low. sbk, the standard error of bk, gives us an indication of how much the
point estimate is likely to vary from the corresponding population parameter.
Let H = the set of all the X (independent) variables.
Let Gk = the set of all the X variables except Xk.
The following formulas then hold:

Review of Multiple Regression — Page 5

General case: 1 − RYH 2
sy This formula makes it clear how
sbk = * standard errors are related to N,
(1 − R X k Gk ) * ( N − K − 1) s X k
2

K, R2, and to the inter-

correlations of the IVs.
2 IV case 1 − RY212 sy When there are only 2 IVs,
s bk = * R2XkGk = R212.
(1 − R122 ) * ( N − K − 1) s X k

1 IV case 1 − R2 s When there is only 1 IV, R2XkGk

sb = * Y = 0.
( N − K − 1) s X

Question: What happens to the standard errors as R2 increases? As N increases? As K

increases? As the multiple correlation between one DV and the others increases?
From the above formulas, it is apparent that
• The bigger R2YH is, the smaller the standard error will be.
• The bigger R2XkGk is (i.e. the more highly correlated Xk is with the other IVs in the model),
the bigger the standard error will be. Indeed, if Xk is perfectly correlated with the other IVs,
the standard error will equal infinity. This is referred to as the problem of multicollinearity,
which will be discussed more later. The problem is that, as the Xs become more highly
correlated, it becomes more and more difficult to determine which X is actually producing
the effect on Y.

Also, recall that 1 - R2XkGk is referred to as the Tolerance of Xk. A tolerance close to 1 means
there is little multicollinearity, whereas a value close to 0 suggests that multicollinearity may
be a threat. The reciprocal of the tolerance is known as the Variance Inflation Factor (VIF).
The VIF shows us how much the variance of the coefficient estimate is being inflated by
multicollinearity.
• Larger sample sizes decrease standard errors (because the denominator gets bigger). This
reflects the fact that larger samples will produce more precise estimates of regression
coefficients.
• Adding more variables to the equation can increase the size of standard errors, especially if
the extra variables do not produce increases in R2. Adding more variables decreases the (N -
K - 1) part of the denominator. More variables can also decrease the tolerance of the variable
and hence increase the standard error. In short, adding extraneous variables to a model tends
to reduce the precision of all your estimates.

Hypothesis Testing. With the above information from the sample data, we can test
hypotheses concerning the population parameters. Remember that hypotheses can be one-tailed
or two-tailed, e.g.
H0: β1 = 0 or H0: β1 = 0
HA: β1 ≠ 0 HA: β1 > 0

Review of Multiple Regression — Page 6

The first is an example of a two-tailed alternative. Sufficiently large positive or negative values
of b1 will lead to rejection of the null hypothesis. The second is an example of a 1-tailed
alternative. In this case, we will only reject the null hypothesis if b1 is sufficiently large and
positive. If b1 is negative, we automatically know that the null hypothesis should not be rejected,
and there is no need to even bother computing the values for the test statistics. You only reject
the null hypothesis if the alternative is better.

EXAMPLE:
H0: β1 = 0
HA: β1 > 0
N = 1000, b1 = -10, t1 = -50. Should you reject the Null? (HINT: Most people say reject. Most
people are wrong. Explain why.)

Hypothesis testing procedure. Although we may not be explicit about it, we should always
go through the following 5 steps when testing hypotheses.
1. Specify the null and alternative hypotheses.
2. Specify the appropriate test statistic. In the case of regression, we typically use a T or F
statistic. For other statistical techniques, a chi-square statistic is often appropriate.
3. Determine the acceptance rejection. What values of the test statistic will lead you to reject or
to not reject the null hypothesis?

Note that the above three steps can, and ideally should be, done before the data have actually
been looked at.
4. Compute the value of the test statistic.
5. Decide whether to accept or reject the null hypothesis. This should be an “automatic”
decision that follows from the results of the previous steps.
With regression, we are commonly interested in the following sorts of hypotheses:
Tests about a single coefficient. To test hypotheses such as
H0: β1 = 0 or H0: β1 = 0
HA: β1 ≠ 0 HA: β1 > 0
we typically use a T-test. The T statistic is computed as
bk − β k 0 bk
TN − K −1 = =
sbk sbk

The latter equality holds if we hypothesize that βk = 0. The degrees of freedom for T are N-K-1.
If the T value is large enough in magnitude, we reject the null hypothesis.
If the alternative hypothesis is two-tailed, we also have the option of using confidence intervals
for hypothesis testing. The confidence interval is

Review of Multiple Regression — Page 7

 bk − ( tα * sbk ) ≤ β k ≤ bk + ( tα * sbk )
2 2

If the null hypothesis specifies a value that lies within the confidence interval, we will not reject
the null. If the hypothesized value is outside the confidence interval, we will reject. Suppose,
for example, bk = 2, sbk = 1, and the “critical” value for T is 1.96. In this case, the confidence
interval will range from .04 to 3.96. If the null hypothesis is βk = 0, we will reject the null,
because 0 does not fall within the confidence interval. Conversely, if the null hypothesis is βk =
1, we will not reject the null hypothesis, because 1 falls within the confidence interval.
If the alternative value is two-tailed and there is only one IV, we can also use the F-statistic. In
the one IV case, F = T2.

Global F Test: Tests about all the beta coefficients. We may want to test whether any of the
betas differ from zero, i.e.
H0: β1 = β2 = β3 = ... = βK = 0
HA: At least one β ≠ 0.
This is equivalent to a test of whether ρ2 = 0 (since if all the betas equal 0, ρ2 must equal 0). The
F statistic, with d.f. = K, N-K-1, is appropriate. As noted above, F = MSR/MSE; or equivalently,
R 2 * ( N − K − 1)
F=
(1 − R 2 ) * K
Question: Looking at the last formula, what happens to F as R2 increases? N increases? K
increases? What are the implications of this?
Note that
• Larger R2 produce bigger values of F. That is, the stronger the relationship is between the
DV and the IVs, the bigger F will be.
• Larger sample sizes also tend to produce bigger values of F. The larger the sample, the less
uncertainty there is whether population parameters actually differ from 0. Particularly with a
large sample, it is necessary to determine whether statistically significant results are also
substantively meaningful. Conversely, in a small sample, even large effects may not be
statistically significant.
• If additional variables do not produce large enough increases in R2, then putting them in the
model can actually decrease F. (Note that as K gets bigger, the numerator can get smaller
and the denominator can get bigger.) Hence, if there is too much “junk” in the model, it may
be difficult to detect important effects.
• The F statistic does not tell you which effects are significant, only that at least one of them is.
• In a bivariate regression (and only in a bivariate regression) Global F = T2.

Incremental F Tests about a subset of coefficients. [NOTE: We will talk about incremental
F tests in great detail toward the middle of the course. For now, I will just briefly review what

Review of Multiple Regression — Page 8

they are.] We sometimes wish to test hypotheses concerning a subset of the variables in a model.
For example, suppose a model includes 3 demographic variables (X1, X2, and X3) and 2
personality measures (X4 and X5). We may want to determine whether the personality measures
actually add anything to the model, i.e. we want to test
H0: β4 = β 5 = 0
HA: β4 and/or β5 ≠ 0
Another common example is when we have multi-category qualitative variables (e.g. Religion,
where 1= Catholic, 2 = Protestant, 3 = Other). Here, we compute a set of Dummy Variables
and then test whether the entire set of dummies is statistically significant. (e.g. X4 = 1 if
Catholic, 0 otherwise; X5 = 1 if Protestant, 0 otherwise. Note that there are always at least one
fewer dummy variables than there are categories, and that one category (the “excluded
category”) is coded 0 on all the dummies.) One way to proceed is as follows.
1. Estimate the model with all 5 IVs included. This is known as the unconstrained model.
Retrieve the values for SSE and/or R2 (hereafter referred to as SSEu and R2u.)
2. Estimate the model using only the 3 demographic variables. We refer to this as the
constrained model, because the coefficients for the excluded variables are, in effect,
constrained to be 0. Retrieve the values for SSE and/or R2 (hereafter referred to as SSEc and
R2c).
3. Compute the following:

( SSE c − SSE u ) / J
FJ , N − K −1 =
SSE u /( N − K − 1)
( SSE c − SSE u ) * ( N − K − 1)
=
SSE u * J
( Ru2 − Rc2 ) * ( N − K − 1)
=
(1 − Ru2 ) * J
where J = the number of constraints imposed (in this case, 2) and K = the number of variables in
the unconstrained model (in this case, 5). Put another way, J = the error d.f. for the constrained
model minus the error d.f. for the unconstrained model.
If J = 1, this procedure will lead you to the same conclusions a two-tailed T test would (the
above F will equal the T2 from the unconstrained model. )
If J = K, i.e. all the IVs are excluded in the constrained model, the incremental F and the Global
F become one and the same; that is, the global F is a special case of the incremental F, where in
the constrained model all variables are constrained to have zero effect. You can see this by
noting that, if there are no variables in the model, R2 = 0; also, SSE =SST, since if nothing is
explained then everything is error (recall too that SSR = SST - SSE).

Review of Multiple Regression — Page 9

When you can use incremental F. In order to use the incremental F test, it must be the case
that
• The sample is the same for each model estimated. This assumption might be violated if, say,
missing data in variables used in the unconstrained model caused the unconstrained sample
to be smaller than the constrained sample. You should be careful how missing data is getting
handled in your statistical routines
• One model must be “nested” within the other; that is, one model must be a constrained, or
special case, of the other. For example, if one model contains IVs X1-X5, and another model
contains X1-X3, the latter is a special case of the former, where the constraints imposed are
β4 = β5 = 0. If, however, the second model included X1-X3 and X6, it would not be nested
within the first model and an incremental F test would not be appropriate.
• Other types of constraints can also be tested with an incremental F test. For example, we
might want to test the hypothesis that β1 = β2, i.e. two variables have equal effects. We’ll
discuss such possibilities later.
Other comments
• Constrained and unconstrained are relative terms. An unconstrained model in one analysis
can be the constrained model in another. In reality, every model is “constrained’ in the sense
that more variables could always be added to it.
• Wald tests, which are easily done in Stata and which we will discuss this semester, are an
alternative to incremental F tests.

Dummy Variables
(1) We frequently want to examine the effects of both quantitative and qualitative
independent variables on quantitative dependent variables. Dummy variables provide a means
by which qualitative variables can be included in regression analysis. The procedure for
computing dummy variables is as follows:

(a) Suppose there are L groups. You will compute L-1 dummy variables. For
example, suppose you had two categories for race, white and black. In this example, L = 2, since
you have two groups. Hence, one dummy variable would be computed. If we had 3 groups (for
example, white, black, and other) we would construct 2 dummy variables.

(b) The first group is coded 1 on the first dummy variable. The other L-1 groups are
coded 0. On the second dummy variable (if there is one), the second group is coded 1, and the
other L-1 groups are coded zero. Repeat this procedure for each of the L-1 dummy variables.

(c) Note that, under this procedure, the Lth group is coded 0 on every dummy
variable. We refer to this as the “excluded category.” Another way of thinking of it is as the
“reference group” against which others will be compared.

For example, suppose our categories were white, black, and other, and we wanted white
to be the excluded category. Then,

Review of Multiple Regression — Page 10

 Dummy1 = 1 if black, 0 if other, 0 if white
Dummy2 = 0 if black, 1 if other, 0 if white

Incidentally, note that if we wanted to compute it, Dummy3 = 1 - Dummy1 - Dummy2.

We do not include Dummy3 in our regression models, because if we did, we would run into a
situation of perfect collinearity. (This should make intuitive sense: If we know someone is not
black and not other, then we know that the person is white.)

Also note that, before computing dummies, you may want to combine some categories if
the Ns for one or more categories are very small. For example, you would have near-perfect
multicollinearity if you had a 1000 cases and one of your categories only had a dozen people in
it. In the present case, if there were very few others, you might just want to compute a single
dummy variable that stood for white/nonwhite.

(2) When a single dummy variable has been constructed and included in the equation, a T
test can be used. When there are multiple dummy variables, an incremental F test (discussed in
more detail elsewhere) is appropriate. Basically, the strategy is to estimate a “constrained”
model that does not include the dummy variables. Then, run a second “unconstrained” model
which has the same variables as the first one plus the dummy variables. You then do an
incremental F test to see whether including the dummy variables significantly increases R2. If
the increase is significant, you conclude that the original categorical variable (race, religion, or
whatever) significantly affects the dependent variable.

(3) EXAMPLE: The dependent variable is income, coded in thousands of dollars. BLACK
= 1 if black, 0 otherwise; OTHER = 1 if other, 0 otherwise. White is the “excluded” category,
and whites are coded 0 on both BLACK and OTHER.

Variable B

BLACK -10.83
OTHER - 5.12
(Constant) 29.83

For whites, predicted income = 29.83 -(10.83 * 0) - (5.12 * 0) = 29.83.

For blacks, predicted income = 29.83 -(10.83 * 1) - (5.12 * 0) = 19.00.
For others, predicted income = 29.83 -(10.83 * 0) - (5.12 * 1) = 24.71.

In this simple example, the constant is the mean for members of the excluded category, whites.
The coefficients for BLACK and OTHER show how the means of blacks and others differ from
the mean of whites. That is why whites can be thought of as the “reference group”: the dummy
variables show how other groups differ from them.

In a more complicated example, in which there were other independent variables, you can think
of the dummy variable coefficients as representing the average difference between a white, a
black and an other who were otherwise identical, i.e. had the same values on the other IVs.

Review of Multiple Regression — Page 11

Example:

Variable B

BLACK - 4.00
OTHER - 2.10
EDUCATION + 1.50
(Constant) +12.00

This model says that, if a white, a black, and an other all had the same level of education, on
average the black would make $4,000 less than the white, while the other would make on
average $2,100 less than the white. So if, for example, each had 10 years of education,

white: predicted income = 12.00 -(4.00* 0) - (2.10 * 0) + (1.50 * 10) = 27.0

black: predicted income = 12.00 -(4.00* 1) - (2.10 * 0) + (1.50 * 10) = 23.0
other: predicted income = 12.00 -(4.00* 0) - (2.10 * 1) + (1.50 * 10) = 24.9.

As a substantive aside, note that the dummy variable coefficients in this hypothetical example
became much smaller once education was added to the model. This often happens in real world
examples. In this case, a possible explanation might be that much, but not all, of the differences
in mean income between the races reflects the fact that whites tend to have more years of
education than do other racial groups.

(4) An alternative to dummy variable coding is effect coding. Computational procedures are
the same, except that the excluded category is coded -1 on every effect variable. Hence, if our
categories were white, black, and other, the effect variables would be coded as

Effect1 = 1 if black, 0 if other, -1 if white

Effect2 = 0 if black, 1 if other, -1 if white

Dummy variable coding and effect coding yield algebraically equivalent results; that is,
you get the same R5, same F values, etc. The estimates of the ß's differ, but you can easily
convert parameters obtained using dummy variable coding to parameters obtained using effect
coding.

Dummy variable coding is probably most commonly used. However, effect coding
provides a means by which 1-way analysis of variance problems can be addressed using multiple
regression. It can be shown that n-way analysis of variance is merely a special case of multiple
regression analysis, and both fall under the heading of the “general linear model”.

(5) (Optional) Yet another alternative to dummy variable coding is contrast coding. Contrast
coding provides the researcher with the most control and flexibility in specifying the group
comparisons or contrasts that are to be tested. I won’t discuss contrast coding in class but the
readings packet has an optional discussion.

Review of Multiple Regression — Page 12

Partial and Semipartial correlations/ Stepwise regression. The partial correlation
measures the strength of the association between Xi and Y when all other X’s are controlled for.
Semipartial correlations (also called part correlations) indicate the “unique” contribution of a
variable. Specifically, the squared semipartial correlation for a variable tells us how much R2
will decrease if that variable is removed from the regression equation. Some relevant formulas
are
2 2
sr k = sr k , pr 2k = sr k2 = sr k
pr k = 2
1 - RYGk 2 2
1 - RYH + sr k 1 − RYGk 1 − RYH
2
+ sr 2k

srk == bk′ * 1 − RX2 k Gk = bk′ * Tolk , srk2 = RYH

2
− RYG
2
k
= bk′2 * (1 − RX2 k Gk ) = bk′2 * Tolk

That is, to get Xk’s unique contribution to R2, first regress Y on all the X’s. Then regress Y on
all the X’s except Xk. The difference between the R2 values is the squared semipartial
correlation. Or alternatively, the standardized coefficients and the Tolerances can be used to
compute the squared semipartials. Note that, the more “tolerant” a variable is (i.e. the less highly
correlated it is with the other IVs), the greater its unique contribution to R2 will be.
Some alternative formulas that may occasionally come in handy are

Tk * 1 − RYH
2
srk =
N − K −1

Tk
prk =
T + ( N − K − 1)
k
2

Note that the only part of the calculations that will change across X variables is the T value;
therefore the X variable with the largest partial and semipartial correlations will also have the
largest T value (in magnitude).

Recall that
• Once one variable is added or removed from an equation, all the other semipartial
correlations can change. The semipartial correlations only tell you about changes to R2 for
one variable at a time.
• Semipartial correlations are used in Stepwise Regression Procedures, where the computer
(rather than the analyst) decides which variables should go into the final equation. In a
forward stepwise regression, the variable which would add the largest increment to R2 (i.e.
the variable which would have the largest semipartial correlation) is added next (provided it
is statistically significant). In a backwards stepwise regression, the variable which would
produce the smallest decrease in R2 (i.e. the variable with the smallest semipartial
correlation) is dropped next (provided it is not statistically significant.)

Review of Multiple Regression — Page 13